As sovereign AI opens up demand for AI semiconductors in national and public sectors, a market previously centred on hyperscalers, the nature of required chips is also changing. AI chip demand is expanding from an all-GPU focus to a full stack that includes CPUs and memory.

Sovereign AI has emerged as a major factor behind this demand. It refers to a country or organisation’s capability to independently control AI within its own laws and regulations. As regulations such as the European Union’s AI Act and the extraterritorial application of the U.S. CLOUD Act turn reliance on U.S. hyperscalers into a structural risk, countries are accelerating the build-out of their own AI infrastructure. Nvidia CEO Jensen Huang (젠슨 황) said at the GPU Technology Conference in March that countries should directly own data centres, in the same context.

The pace of change is accelerating, in particular with the spread of agentic AI. While token generation is the domain of accelerators such as GPUs, in an environment where agents create agents and operate around the clock, token scheduling, orchestration and service management are handled by central processing units. Arm CEO Rene Haas (르네 하스) said at the launch of a CPU product for AGI that CPU demand in an agentic AI environment would quadruple, adding that this was a conservative estimate. Sovereign AI amplifies that demand at a national level.

The market’s scale is also borne out in numbers. According to Kyobo Securities, Nvidia’s fiscal 2026 sovereign AI revenue more than tripled from a year earlier, topping $30 billion. Global market research firm Gartner forecast the AI cloud market would reach $267 billion in 2030, with sovereign-focused neo-cloud accounting for 20 percent. An analysis has also said countries seeking to build their own AI stacks should invest at least 1 percent of gross domestic product in infrastructure.

From the semiconductor industry’s perspective, the growth of this demand differs from the past. Sovereign environments face major constraints in power, cooling and space. Unlike hyperscale clouds, power efficiency rather than absolute performance, and CPU-memory system integration rather than a single accelerator, rise in priority.

In the industry, the view is spreading that what is needed is not necessarily a cloud-scale system costing several trillion won but infrastructure tailored to required scale. After large language models, models are diversifying into areas such as vision-language models, and a trend is also joining in to run appropriately sized models at the network edge, including base stations and relay stations.

◆ Custom demand spreads beyond GPUs to CPUs, memory and packaging

This is also why Nvidia is focusing not only on GPUs but also on its internally developed custom CPU, Vera Rubin. Arm has also disclosed its own designed server CPU, the AGI CPU, and entered the market head-on. It is targeting replacement of x86, touting twice the performance at the same power. Arm President Mohamed said customers had asked directly for help, saying current solutions were insufficient. With designs optimised for sovereign environments where power and cooling constraints are significant, the spread of sovereign AI is accelerating Arm’s entry into the server chip market.

In South Korea, a structure is also taking shape in which investment in sovereign AI infrastructure leads to chip demand. According to Kyobo Securities, South Korea allocated an AI budget of 9.9 trillion won and in the first phase invested 1.4 trillion won to secure 13,000 GPUs. NHN Cloud, Naver Cloud and Kakao were selected as operators, and NHN Cloud projects revenue of about 300 billion won over 5 years from external supply outside private-sector use. The second phase, worth about 2.08 trillion won, aims to procure additional latest GPUs on the Blackwell to Vera Rubin level and build an always-on GPUaaS system. Full-stack demand is expected to expand alongside GPU procurement, including server CPUs, memory and packaging.

As Gartner forecasts that more than 75 percent of European and Middle Eastern companies will move workloads to in-country solutions by 2030, the trend is a structural shift rather than temporary policy-driven demand. For the chip industry, it can be seen as an opportunity to diversify customers while expanding product portfolios.

As sovereign AI demand grows to hyperscaler-level volumes, the AI semiconductor market shifts from a single GPU axis to a multi-axis structure spanning CPUs, memory and packaging. As it diversifies, characteristics also differ depending on the nature of sovereign AI. An industry official said, "In the end, integration of chips and systems becomes important."